This conference paper was submitted for presentation at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

Plastic components for technical applications are often reinforced with short fibers. For the simulation of such components, the direction-dependent mechanical material properties must be taken into account. This is done by using anisotropic material models. The required model parameters are usually determined using an iterative re-engineering approach by comparing experimentally determined and simulatively predicted test data. For this purpose, the stress-strain curves determined from a tensile test using a specimen with a known fiber orientation are compared with the simulated predicted results and the model parameters of the anisotropic material model are adjusted iteratively until sufficient agreement is achieved. This calibration process is comparatively complex. It requires the execution of an injection molding simulation for the specimen to determine the local fiber orientation, which is manufacturing-dependent. Furthermore, a structural simulation of the tensile test for this test specimen is required. In industrial practice, this calibration process is often limited for various reasons: the responsible employee only has access to a structural simulation program, but not to an injection molding simulation program, lack of expertise or experience to determine the model parameters in a physically meaningful way, time pressure etc. This article presents a method based on machine learning (ML) that makes it possible to carry out the outlined calibration process largely automatically without having to use conventional FEM and injection molding simulation programs. For this purpose, artificial neural networks and decision tree-based ML models are trained for the regression of target variables. In order to generate training data, injection molding and structural simulations are carried out for characteristic plastic material groups and typical specimen geometries. The rheological and mechanical material properties and geometric ratios are varied. The trained ML models can completely replace the numerical simulations. This means that the entire calibration process can be carried out autonomously by the responsible employee. A retained test data set that was not used for training was used to validate the performance of the ML models. Initial results currently available show very good agreement between experimental and predicted stress-strain curves.